furybot

1. Sexual content involving minors

Absolute, in any form, real or synthetic, including stylised and animated depictions. Required by 18 U.S.C. §§ 2251–2252A and § 1466A, which reach drawings and computer-generated images as well as photographs. This limit has its own vocabulary, its own handling of deliberate obfuscation and its own test suite, and uploaded images are additionally checked by a scanner running on infrastructure we control before they reach a model. It is the one place where we accept a wrongly refused request as the price of not getting this wrong.

2. Sexual imagery of a real, identifiable person

Non-consensual intimate imagery, required by the TAKE IT DOWN Act, the civil remedy at 15 U.S.C. § 6851 and most states' NCII statutes. Both signals have to be present, which is deliberate: a request that is merely sexual is a category question, and a photograph of a named person is entirely ordinary. It is the combination that is unlawful, and a label does not cure it, so this is a refusal rather than a disclosure.

3. Deceptive content about a real person

Refused where the evident purpose is to deceive about a real, identifiable individual in a way that causes harm — fabricated evidence, a forged document, a false endorsement, a statement they never made. State election-deepfake statutes, the FTC Act on deceptive endorsements, and ordinary defamation and fraud law. Note where this boundary sits, because a great deal of legitimate work lives right beside it: satire, caricature, parody and political commentary are permitted. A cartoon of a president is not deception. A photorealistic fake of that president appearing to commit a crime, presented as real, is.

4. Material that would facilitate serious harm

Refused where content would provide meaningful assistance toward mass casualties or serious violent crime. This is about operational usefulness, not subject matter — depicting violence, war, weapons or crime is not blocked. Fiction, history and journalism are not covered by it: a war film scene is not a threat.

Labelled, not refused: synthetic images of real people

A synthetic image of a real, identifiable person is allowed and carries an AI-generated label. Whether such an image is satire, comment or deception is not a judgement we make — but a viewer is entitled to know it was generated, and disclosure is what the relevant statutes ask for.

Politics is not on the list

Political subjects are permitted by default and are not filtered by viewpoint. Satire, protest imagery, campaign parody and criticism of any government, party, movement or figure are ordinary requests here, in every direction. This is not merely a promise: the test suite pairs mirrored prompts from opposing sides of political, religious and identity debates and fails the build unless each pair is treated identically — so a term added to one side later cannot slip in unnoticed.

One honest caveat about the chat models

The chat models are third-party open-weight models, and a model carries the biases of its training however our policy is written. We add no political layer of our own, and the refusal logic above is ours and is symmetric — but we cannot remove a lean that is baked into a model's weights. The current models are known to push back on criticism of the Chinese government, which is a property of the weights and not of anything we do. We would rather name that than let you assume a neutrality we cannot deliver.

Adult content is absent, not disabled

Sexual content and nudity cannot be produced here, and that is not a switch in the off position — the capability is not present in the software we run, and there is no configuration, administrator action or support request that introduces it. This is a business decision rather than a moral one. It is what lets an ordinary payment processor work with us, and it is why the rest of the product can be as permissive as it is.

Everything else is a setting, not a decision we made

Mature, graphic, political and otherwise difficult subject matter is governed by the profile on your account rather than by us. Set it strictly for a shared or family machine — the Strict profile excludes political subjects along with violence and other mature themes — or open it up for professional work that needs the latitude. You can change it at any time, and the page you are reading tells you exactly where the fixed limits are.

Common questions

Can I turn off the four limits?

No. All four are legal requirements and they are compiled into the software rather than configured. There is no setting, administrator action or support request that lifts any of them.

Is the tool politically biased?

The refusal logic is tested for symmetry directly: mirrored prompts from opposing sides of political, religious and identity debates must receive identical verdicts, and the test suite fails the build if they do not. The chat models are a separate matter — they are third-party open-weight models and carry their own training biases, which we disclose rather than paper over.

Can I make political memes and satire?

Yes, in any direction, about anyone. Satire, caricature and political commentary are explicitly permitted. The line is deception presented as real, not criticism.

Who decides what counts as blocked?

The list above is the whole of it. Everything else is a setting on your account.